针对青藏高原跨境珍稀动物检测中样本不平衡、生境异质性及计算效率等问题,在YOLOv11n基础上提出改进模型YOLOv11-MS。该模型通过引入SlideLoss损失函数缓解样本不平衡以提升泛化能力,采用Slim-neck架构精简特征融合链路减少计算冗余,并设计多尺度扩张注意力（MSDA）模块增强局部细节与全局上下文捕捉能力,有效适配裸岩、湿地等异质生境。在自制的青藏高原跨境珍稀动物数据集（含6 921张图像、4类物种）上训练后,模型mAP@0.5达97.3%,推理时间0.9 ms,计算量5.8GFLOPs,在精度与效率间实现平衡。研究成果可为跨境生态合作及“一带一路”绿色发展倡议提供技术支撑。
Abstract
To address issues such as class imbalance, habitat heterogeneity, and insufficient computational efficiency in the detection of cross-border rare animals on the Qinghai-Tibet Plateau, this paper proposes an improved model named YOLOv11-MS based on YOLOv11n: the model introduces the SlideLoss loss function to alleviate class imbalance and enhance generalization ability, adopts the Slim-neck architecture to streamline the feature fusion pipeline and reduce computational redundancy, and designs a multi-scale dilated attention (MSDA) module to strengthen the capture of local details and global context, effectively adapting to heterogeneous habitats like bare rocks and wetlands. Trained on a self-constructed cross-border rare animal dataset of the Qinghai-Tibet Plateau (containing 6,921 images and 4 species), the model achieves a mAP@0.5 of 97.3%, an inference time of 0.9 ms, and a computational complexity of 5.8 GFLOPs, striking a balance between accuracy and efficiency, and its research outcomes can provide technical support for cross-border ecological cooperation and the Belt and Road Green Development Initiative.
关键词
YOLOv11-MS /
青藏高原 /
跨境珍稀动物 /
目标检测 /
深度学习
Key words
YOLOv11-MS /
Qinghai-Tibet Plateau /
cross-border rare animals /
object detection /
deep learning
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] 应凌霄,张路,张力,等.青藏高原生物多样性保护面临的问题与挑战[J].国家公园(中英文),2025,3(5):338-348.
[2] 蒋昆财,喻钧,胡志毅,等.YOLOv10的SAR图像海面目标检测方法研究[J].西安工业大学学报,2025,45(3):468-478.
[2] 王俊康. 基于改进的YOLOv7“低慢小”目标识别分析[J].数字技术与应用,2025,43(6):191-194.
[4] 张海宁. 基于YOLO算法的模糊场景人车检测研究[D].大连:大连交通大学,2025.
[5] 洪成坤,杨涛,付丽媛.基于改进YOLOv5算法的脑出血CT图像识别与分割研究[J].医疗卫生装备,2025,46(5):1-8.
[6] Zhou K,Jiang S.Forest fire detection algorithm based on improved YOLOv11n[J].Sensors,2025,25(10):2989.
[7] Jiao J,Tang Y M,Lin K Y,et al.Dilateformer:Multi-scale dilated transformer for visual recognition[J].IEEE transactions on multimedia, 2023,25:8906-8919.
[8] Li H,Li J,Wei H,et al.Slim-neck by GSConv:A lightweight-design for real-time detector architectures[J].Journal of Real-Time Image Processing,2024,21(3):62.
[9] Yu Z,Huang H,Chen W,et al.Yolo-facev2:A scale and occlusion aware face detector[J].Pattern Recognition,2024,155:110714.
[10] Yu Y,Khadivi S,Xu J.Can data diversity enhance learning generalization?[C]//Proceedings of the 29th international conference on computational linguistics,2022:4933-4945.
[11] Kellenberger B,Marcos D,Tuia D.Detecting mammals in UAV images:Best practices to address a substantially imbalanced dataset with deep learning[J].Remote sensing of environment,2018,216:139-153.
[12] Pressey R L,Logan V S.Level of geographical subdivision and its effects on assessments of reserve coverage: a review of regional studies[J].Conservation Biology,1994,8(4):1037-1046.
[13] Chen L,Li G,Zhang S,et al.YOLO-SAG:An improved wildlife object detection algorithm based on YOLOv8n[J].Ecological Informatics,2024,83:102791.
[14] Pommé L E,Bourqui R,Giot R,et al.Relative confusion matrix:Efficient comparison of decision models[C]//2022 26th International Conference Information Visualisation (IV).IEEE,2022:98-103.
[15] Wilson C L,Blue J L,Omidvar O M.Training dynamics and neural network performance[J].Neural Networks,1997,10(5):907-923.
[16] Shaham U,Yamada Y,Negahban S.Understanding adversarial training:Increasing local stability of supervised models through robust optimization[J].Neurocomputing,2018,307:195-204.
